Perceiver IO added to HuggingFace Transformers for multi-modal inputs

AI Impact Summary

Perceiver IO extends the Transformer family to natively handle text, images, audio, video, and other modalities by performing cross-attention with a latent set of tokens. This approach eliminates the quadratic compute bottleneck of full self-attention on large inputs, enabling scalable multi-modal inference in HuggingFace Transformers via PerceiverModel and its pre/post processors. For engineering teams, this provides a single, extensible path (using PerceiverTokenizer, PerceiverTextPreprocessor, PerceiverClassificationDecoder) to deploy multi-modal pipelines, potentially lowering latency and upkeep compared to modality-specific architectures.